Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.deck.co/llms.txt

Use this file to discover all available pages before exploring further.

Deck provides access to computer use agents and the infrastructure that runs them. Computer use agents operate websites and software like a human would, giving you access and automation for services that lack APIs to integrate against.

Architecture

Every operation follows this pattern. Your application talks to the REST API. The API validates the request, queues the work, and delegates to a managed agent. The agent does the work and reports back through the API and event deliveries.

Compute orchestration

Deck provisions isolated sessions on demand. When you run a task, Deck spins up a session, optionally authenticates using stored credentials, and executes the required actions.

Session management

A session is created automatically when a task runs. Deck connects to the source, optionally authenticates using the credential, executes the work, and cleans up the session when done. You can reuse a session across multiple task runs by passing the session_id, which avoids re-authentication and is faster. Sessions are cleaned up automatically after the last task run completes or after a period of inactivity.

Execution model

All work in Deck is asynchronous. When you call the API to run a task, you get an immediate response with a queued status. The actual work happens in the background.
1

Request

Your application calls the Deck API with the required parameters. Deck validates the input against the task’s schema and rejects malformed requests synchronously.
2

Execute

Deck assigns an agent, which navigates the target website and performs the defined actions: clicking, typing, reading, downloading.
3

Return

The agent extracts data according to the task’s output schema. Deck validates the output, stores it on the task run, and emits a completion event.

Deterministic output

Every task defines an input schema and an output schema. Regardless of how the underlying website renders its data, Deck returns a consistent JSON structure. The output schema guides what the agent extracts, but fields may be null if the data isn’t available on the source. This contract means you write your integration code once and it works across sources. A “fetch reservations” task returns the same shape whether the source is Hilton, Marriott, or Hyatt.

Handling interactions

Some operations require human input midway through execution. MFA codes, security questions, and other verification prompts. The agent can’t proceed without your user’s involvement. When this happens, Deck pauses the agent and notifies your application:
  1. The agent encounters a prompt it can’t resolve autonomously
  2. Deck sets the status to interaction_required and emits an event
  3. Your application collects the required input from the user
  4. You submit the input via the interaction endpoint
  5. The agent resumes exactly where it left off
The session stays open while waiting.

Event-driven integration

Deck is built for event-driven architectures. Rather than polling for status changes, you register event destinations and Deck pushes events to your infrastructure as they happen. Events cover the full lifecycle: credential verification, sessions starting and completing, task runs progressing, interactions required, deliveries failing. Every event includes a summary of the resource so your handler has the context it needs without additional API calls. Deck delivers events to webhooks, AWS SQS, GCP Pub/Sub, and other cloud-native destinations, with built-in retries and delivery tracking.

Next steps

Quickstart

Create an agent, store credentials, and run your first task.

Concepts

Agents, sources, credentials, sessions, tasks, and how they fit together.

Building your auth flow

Collect credentials and manage credential state.

Using the API

Authentication, pagination, idempotency, and error handling.